From Failure Taxonomy to Intervention: A Diagnostic Methodology for Industry-Scale AVLM in Video and Live-Streaming Platform Moderation
This work creates a diagnostic system for large Audio-Visual-Language Models (AVLM) by identifying specific failure modes in their training pipeline. It maps these failures to actionable interventions, such as better data sampling and cross-modal supervision. The goal is to improve the model's general audio-visual reasoning and its performance on real-world content moderation tasks.